Comparison study on time series forecasting techniques for apparel retailing

Abstract

Sales forecasting is the foundation for planning various phases of a firm's business operations. It is also crucial to dynamic supply chains and greatly affects retailers and other channel members in various ways. Effective sales forecasting enables big improvement in supply chain performance. In today's apparel retailing, sales forecasting mainly rely on subjective assessment and experience of sales/marketing personnel with simple statistical analysis of historical sales data. While there exist various sales forecasting techniques, it is unknown how each technique fits different types of apparel sales data, and no research efforts have ever been made to investigate and compare the effects of different techniques on different sales data patterns for apparel retailing. The purpose of this research is to investigate and compare the forecasting performances of commonly used univariate and multivariate time series forecasting techniques for apparel retailing. A methodology with nine procedures was presented to compare different time series forecasting techniques for apparel retailing. Four typical data patterns of apparel retailing were identified to represent various apparel sales time series. Some commonly used time series forecasting techniques, including five univariate techniques, three multivariate techniques and neural network (NN) techniques, were used and compared. Five accuracy measures were used to evaluate the forecasting results, which included the mean absolute deviation, mean absolute error, mean absolute percentage error, mean absolute scaled error and root mean square error. Five commonly used univariate forecasting techniques were used to construct nine forecasting models. The performances of these univariate forecasting models and two univariate NN models were compared on the basis of a large number of apparel sales time series. These apparel sales data were collected from an apparel retail company and categorized into trend, seasonal, irregular and random patterns. 10 multivariate forecasting models were constructed based on four multivariate forecasting techniques. The forecasting performances of these models and two multivariate NN (MVNN) models were compared on the basis of the same apparel sales data. Lastly, the performances generated by these univariate models were further compared with those by the multivariate models. This research also investigated the effects of different numbers of input variables and different accuracy measures on sales forecasting performances. The comparison study showed that (1) for different data patterns, forecasting performances generated by univariate and multivariate forecasting models are mixed; for seasonal data patterns, ARX(3,2), ARMAX(3,3,2) and NN(3) models can perform better than the others; for irregular data patterns, ARMAX(3,3,2), ARMAX(3,3,1) and ARMAX(2,2,2) models can perform better than the others; for random data patterns, AR(2), ARX(3,1), ARMA(1,1) and ARMAX(3,3,1) models can outperform the other models. (2) Among the univariate techniques, the moving average technique usually generates the worse forecasting results no matter what data pattern is used. (3) NN models cannot provide better forecasting performances than the other classical models. (4) The multivariate time series forecasting models cannot always generate better results than the univariate time series forecasting models. For example, the AR(2,1) and AR(2,2) models usually cannot generate better forecasts than the AR(2) models although the ARX(3,1) and ARX(3,2) models are relatively better than the AR(3) model. (5) The MVNN models cannot outperform the other traditional multivariate models significantly. (6) Even for the same model, different parameter settings can impact forecasting results greatly. For instance, the ARX(3,2) model generates much better results than the ARX(2,2) model for seasonal patterns. (7) In addition, different accuracy measures and different numbers of input variables can impact forecasting results greatly. These comparison results show that it is important to select appropriate forecasting models based on different data patterns, and to set appropriate model parameters, NN structures, accuracy measures and input variables based on specific forecasting tasks. The comparison presented in this thesis can provide a theoretical basis for researchers and practitioners of apparel sales forecasting, and help them select the appropriate forecasting or benchmark models for different apparel sales forecasting tasks.